AI-Based Segmentation Reduces the Retrospective Miss Rate of Pancreatic Ductal Adenocarcinoma.
Authors
Affiliations (5)
Affiliations (5)
- Joint Department of Medical Imaging, Sinai Health System, Princess Margaret Hospital Cancer, University of Toronto, 600 University Ave. Rm 2-220, Toronto, M5G1X5, Canada.
- Ontario Institute for Cancer Research, Toronto, ON, Canada.
- Wallace McCain Centre for Pancreatic Cancer, Princess Margaret Hospital Cancer Centre, Toronto, ON, Canada.
- Joint Department of Medical Imaging, Sinai Health System, Princess Margaret Hospital Cancer, University of Toronto, 600 University Ave. Rm 2-220, Toronto, M5G1X5, Canada. [email protected].
- Lunenfeld Tanenbaum Research Institue, Sinai Health System, Toronto, Canada. [email protected].
Abstract
The objective of this study is to evaluate the performance of an nnU-Net v2-based artificial intelligence (AI) model trained to segment the pancreas anatomy and lesions when present in reducing the retrospective miss rate (RMR) of pancreatic ductal adenocarcinoma (PDAC) on CT. In this single-center retrospective study, 132 patients with pathology-proven PDAC and prior contrast-enhanced CT (2011-2022) were included alongside 80 public-domain controls. An nnU-Net v2 model was trained to segment PDAC on 683 independent cases (501 public, 182 internal) to segment the pancreas anatomy and to identify and segment lesions. RMR was defined as the proportion of retrospectively visible but unreported lesions. Model performance was assessed for direct lesion detection and for lesion plus indirect signs (ductal dilatation/atrophy). McNemar's test was used for paired comparisons, and diagnostic metrics including sensitivity, specificity, PPV, NPV, AUC, and Youden's J were calculated with 95% CIs. The radiologist RMR was 33.3% (44/132). AI reduced this to 23.5% (31/132), and combined AI + radiologist interpretation achieved 16.6% (22/132) (p = 0.0196). For tumors ≥ 2 cm at diagnosis, RMR decreased from 36 to 13% with combined reading (p < 0.001). Including indirect signs, AI achieved an RMR of 12% and combined reading 10%. Sensitivity and specificity were 77% and 96%, respectively, improving to 88% sensitivity when indirect signs were included; the false-positive rate was 3.8%. An nnU-Net v2 model significantly reduced the RMR of PDAC on CT, particularly for larger tumors and when indirect signs were incorporated, with a low false-positive rate.